This session covers how large language models can be used to support complex, high-impact decisions when data and evidence are spread across many documents, systems, and scientific sources. In many real situations, the challenge is not a lack of information, but the effort required to find relevant evidence, connect it across sources, identify gaps or conflicting information, and make sense of it within a limited amount of time.

We use federal Earth science and Earth-observation scenarios to make the problem concrete. The examples draw on scientific and program information from NASA, NOAA, USGS, EPA, and related federal sources, including research solicitations and award portfolios, satellite and remote-sensing products, technical reports, policy guidance, operational monitoring information, and funding records. The scenarios include AI/ML research and funding analysis, scientific portfolio overlap, operational readiness, uncertainty, and the use of Earth-observation missions and sensors such as NASA PACE, Landsat, Sentinel-2, MODIS, VIIRS, and Sentinel-3. Water-quality remote sensing provides one practical example, but the approach is applicable more broadly to Earth science, satellite imagery, AI/ML-enabled research, scientific program planning, research funding, and other high-stakes federal decision-support problems.

We focus on a practical approach called agentic Retrieval-Augmented Generation (RAG). Instead of relying on a single model to answer questions, we describe systems where multiple LLM-based agents work together. One agent gathers relevant material, another checks evidence across sources, and another pulls the information together while highlighting uncertainty. Topics include keeping model outputs grounded in real data, dealing with inconsistent information, evaluating source relevance and recency, identifying possible overlap or gaps across funded research, and understanding where models can fail. The same framework can help connect scientific evidence with program priorities, funding history, operational needs, and mission-specific constraints.

The goal of this talk is not automation, but better decision support. Agentic RAG can help analysts and decision makers spend less time manually searching across scientific reports, satellite and Earth-observation information, research portfolios, funding records, policy documents, and operational data, and more time evaluating evidence and making informed decisions. Attendees will leave with a clear, realistic picture of how agentic RAG can support high-impact scientific and program decisions while keeping human experts responsible for final judgment.

Target audience: Data scientists, AI/ML practitioners, Earth and environmental scientists, federal program and research analysts, technical leaders, and professionals interested in responsible LLM/RAG-based decision support.
Preferred session duration: 30–45 minutes, with additional time for Q&A.
Technical requirements: None.
Additional information: The session uses practical federal Earth science and Earth-observation examples based on publicly available NASA, NOAA, USGS, and EPA materials, including research portfolios, funding information, scientific reports, and satellite missions/sensors such as PACE, Landsat, Sentinel-2, MODIS, VIIRS, and Sentinel-3. No advanced knowledge of RAG is required; basic familiarity with AI/ML is helpful.